Modal based traditional Structural Health Monitoring techniques are limited because of several factors – including a poorly-formed aggregate system model, very low SNR, and unrealistic boundary conditions. Moreover, global techniques often rely on modal damage indicators that are not sensitive to localized damage.
In this dissertation, the author proposes a new Damage Detection technique that addresses the space-frequency localization of damage artifacts in a reference and no-reference framework. For the first situation of referenced damage detection, the author employs the use of compactly supported subband space/frequency and time/frequency analysis using local vibration characteristics, overcoming the signal to noise ration problem with a nearfield adaptive beamformer filter bank. The beamformer filter bank operates on the subband space and provides accurate spatial selectivity and high signal to noise ratio for any given scan direction. Subband analysis is performed using wavelet packets and Daubechies mother wavelets. The system is simulated using a one dimensional Finite Element model of a simply supported beam with simple constraints as a good approximation of a real situation. The local damage is simulated as a reduction of the Young’s modulus over a selected group of elements.
The Damage Detection is performed using as a damage feature the subband energy for any given scan direction and for each subband center frequency. The energy signature for every location/frequency is compared to the energy signature obtained for the equivalent undamaged structure. The obtained results are validated against the analysis obtained before the beamforming stage, and the algorithm localizes the damage in areas of high probability around the direction of the simulated discontinuity. Moreover, the proposed technique shows a very high accuracy and it is able to detect variations on the structure parameters as low as 1%, with a signal near the noise level.
For the second situation of Damage Detection performed without an undamaged reference for the analysis, the author proposes a new statistical method based on the density estimation of the vibration signal. This technique is based in the Gaussian Mixture estimation of the probability density function of the vibration signal, using a greedy EM approach with a new model order selection criteria. This model order is based on global measurement on the cumulative density function as well as on local measurement on density indicators, such as the Kullback-Leibler divergence and the estimated Correlation Coefficient. The technique is used to estimate the density of time domain signal and frequency domain signal. As damage indicators, the technique uses the first two principal components from measurements of standard deviation, kurtosis, skewness and entropy on the estimated density. The obtained damage indicators perform better in frequency domain and damage as low as 30% can be detected in a noisy environment.